Abstract
Indoor localization is of particular interest due to its immense practical applications. However, the rich multipath and high penetration loss of indoor wireless signal propagation make this task arduous. Though recently studied fingerprint-based techniques can handle the multipath effects, the sensitivity of the localization performance to channel fluctuation is a drawback. To address the latter challenge, we adopt an artificial multi-layer neural network (MNN) to learn the complex channel impulse responses (CIRs) as fingerprint measurements. However, the performance of the location classification using MNN critically depends on the correlation among the training data. Therefore, we design two different decorrelation filters that preprocess the training data for discriminative learning. The first one is a linear whitening filter combined with the principal component analysis (PCA), which forces the covariance matrix of different feature dimensions to be identity. The other filter is a nonlinear quantizer that is optimized to minimize the distortion incurred by the quantization. Numerical results using indoor channel models illustrate the significant improvement of the proposed decorrelation MNN (DMNN) compared to other benchmarks.
Introduction
While Global Positioning System (GPS) technology has revolutionized outdoor navigation, it remains largely ineffective indoors due to signal blockage and complex multipath effects. In indoor environments, signals bounce off walls, furniture, and other obstacles, creating a chaotic array of echoes. To overcome this, researchers utilize fingerprinting techniques, which record a unique "Channel Impulse Response" (CIR) for specific locations, acting as a digital signature.
A significant challenge in fingerprint-based localization is that fingerprints from adjacent areas, such as neighboring offices, often appear highly correlated. When this data is fed into a standard Multi-layer Neural Network (MNN), the model can become "entangled," where small signal fluctuations lead to incorrect predictions. This paper introduces the Decorrelation MNN (DMNN) framework to address these issues by preprocessing data to ensure distinct, separable features.
Methodology: The DMNN Framework
The core innovation of this work is the introduction of a preprocessing stage before the neural network, utilizing one of two decorrelation filters to "clean" the input data.
1. Linear Whitening Filter (WMNN)
This method employs Principal Component Analysis (PCA) followed by a whitening transformation. By rotating and scaling the data, the filter ensures the covariance matrix of the features becomes an identity matrix. This process makes every feature independent and equally weighted, preventing the network from over-relying on redundant signal components.
2. Nonlinear Optimal Quantizer (QMNN)
Unlike standard compression, this filter maps continuous signal values onto a discrete grid. Using the Newton-Raphson method, the quantizer is optimized to minimize distortion while forcing correlated signal data into distinct "buckets." This acts as a robust feature extractor, filtering out the "jitter" that typically causes mis-classification between adjacent rooms.
Following decorrelation, the data is processed by a 4-layer MNN for final location classification.
Experimental Setup
The proposed system was evaluated using numerical simulations based on the ITU indoor office channel model at 3.5 GHz.
The simulated environment consisted of a 60m x 9m office space divided into 15 to 30 equal-sized rooms. Localization was performed using 3 to 5 distributed single-antenna sensors to collect CIR fingerprints. The training set included 100 measurements per location at a Signal-to-Noise Ratio (SNR) of 20 dB.
Results and Discussion
The DMNN variants demonstrated significant performance gains over traditional benchmarks such as Support Vector Machines (SVM) and standard MNNs.
- Accuracy: At 20 dB SNR, both WMNN and QMNN achieved a mis-classification rate as low as 3–4%. In contrast, standard models exhibited much higher error rates, highlighting the necessity of the decorrelation step.
- Scalability: Increasing the environment's complexity (e.g., doubling the number of offices to 30) initially raised the error rate to 12% with 3 sensors. However, by increasing spatial diversity to 5 sensors, the error rate was successfully reduced back to the 3–4% range.
- Robustness: The study reveals that whitening acts as an effective regularizer, while quantization serves as a powerful tool for noise reduction and feature separation in highly correlated physical signal environments.
Conclusion and Future Work
The DMNN framework offers a promising solution for high-accuracy indoor localization by mitigating the "entanglement" trap of correlated signal data. However, several areas remain for future exploration. The current study assumes perfect CIR knowledge; real-world estimation noise may impact filter performance. Additionally, while the simulation accounts for small-scale fading, dynamic environmental changes (e.g., moving obstacles) require further study. Finally, the computational overhead of real-time PCA and whitening on low-power devices must be addressed to ensure practical deployment.
Summary updated